Credential Error in Microsoft Agent Framework SDK

Hi Yi Chun Lam

agent_framework.exceptions.ServiceInitializationError: Azure credential is required when agents_client is not provided. happens when the client cannot find a usable Azure credential object (it’s None / not passed / not valid), and you didn’t pass a pre-built agents_client.

So the SDK has no way to authenticate to Azure AI Agents / Azure AI Project. [agent_fram...soft Learn | Learn.Microsoft.com], [azure.ai.a...soft Learn | Learn.Microsoft.com], [agent_fram...soft Learn | Learn.Microsoft.com]

In most lab environments, your code creates a credential like DefaultAzureCredential() or AzureCliCredential(), but you are not logged in (or the credential chain can’t find any valid auth method). Then the SDK effectively behaves like “no credential provided”, triggering this error. The official quickstart explicitly calls out using Azure CLI auth (az login) and having access to the Azure AI project. [Microsoft...soft Learn | Learn.Microsoft.com], [Create and...soft Learn | Learn.Microsoft.com]

3) Fix #1 (recommended for labs): use Azure CLI credential and ensure you’re logged in

Step A — Login

Run this in your terminal:

Shell

az login
az account show

This ensures the CLI token exists for AzureCliCredential() to use. [Microsoft...soft Learn | Learn.Microsoft.com], [Create and...soft Learn | Learn.Microsoft.com]

4)Step B — Use the credential correctly in code

The AzureAIAgentClient expects a credential= parameter (per the SDK docs). If you pass async_credential= to AzureAIAgentClient, it may not bind to the right parameter and the SDK will treat it like no credential was provided. Use credential=... like below. [agent_fram...soft Learn | Learn.Microsoft.com], [Microsoft...soft Learn | Learn.Microsoft.com]

import asyncio
from azure.identity.aio import AzureCliCredential
from agent_framework.azure import AzureAIAgentClient
async def main():
async with AzureCliCredential() as credential:
async with AzureAIAgentClient(credential=credential) as client:
# run your agent calls here
pass
if __name__ == "__main__":
asyncio.run(main())

This matches the documented constructor usage for AzureAIAgentClient(... credential= ...). [agent_fram...soft Learn | Learn.Microsoft.com], [Microsoft...soft Learn | Learn.Microsoft.com]

5. Fix #2: Make sure required environment variables are set (very common missing step)

Even with a valid credential, the SDK also needs to know where your Azure AI project is and which model deployment to use.

Set these (either in your shell or in a .env file):

• AZURE_AI_PROJECT_ENDPOINT
• AZURE_AI_MODEL_DEPLOYMENT_NAME

The SDK docs and quickstart show these exact variables. [agent_fram...soft Learn | Learn.Microsoft.com], [Microsoft...soft Learn | Learn.Microsoft.com]

Example:

Shell

export AZURE_AI_PROJECT_ENDPOINT="https://<your-project>.cognitiveservices.azure.com"
export AZURE_AI_MODEL_DEPLOYMENT_NAME="gpt-4o-mini"

Or load from .env:

Python

client = AzureAIAgentClient(credential=credential, env_file_path=".env")

[agent_fram...soft Learn | Learn.Microsoft.com], [Microsoft...soft Learn | Learn.Microsoft.com]

Fix #3: If you really intended to pass a pre-built AgentsClient

The error message says “credential is required when agents_client is not provided”. That means alternative path is: create an AgentsClient(endpoint, credential) yourself and pass it in as agents_client=.... The AgentsClient constructor requires both endpoint and credential. [azure.ai.a...soft Learn | Learn.Microsoft.com]

Conceptually:

Python

from azure.ai.agents.aio import AgentsClient
from azure.identity.aio import AzureCliCredential
from agent_framework.azure import AzureAIAgentClient
agents = AgentsClient(endpoint=PROJECT_ENDPOINT, credential=credential)
client = AzureAIAgentClient(agents_client=agents, credential=credential)

(Exact import path may differ by version, but the key is: AgentsClient requires endpoint + credential.) [agent_fram...soft Learn | Learn.Microsoft.com]

6) Quick checklist (this resolves 95% of cases)

1. pip show agent-framework agent-framework-azure-ai azure-identity (packages installed)
2. az login completed successfully
3. You used AzureCliCredential() or a working DefaultAzureCredential()
4. You passed it as credential=credential into AzureAIAgentClient
5. AZURE_AI_PROJECT_ENDPOINT + AZURE_AI_MODEL_DEPLOYMENT_NAME are set

These align directly with the SDK constructor + quickstart requirements. [agent_fram...soft Learn | Learn.Microsoft.com], [Microsoft...soft Learn | Learn.Microsoft.com], [Create and...soft Learn | Learn.Microsoft.com]

Please let me know and if there are any remaining questions or additional details, I can help with, I’ll be glad to provide further clarification or guidance.

Thankyou!

I had a look at AzureAIAgentClient docs here and I found that they wrote the wrong credential parameter and they missed to specify env_file_path.

I was able to run successfully the lab with the following code:

```ruby
# Create a chat agent
async with (
    AzureCliCredential() as credential,
    ChatAgent(
        chat_client=AzureAIAgentClient(credential=credential, env_file_path=".env"),
        name="expenses_agent",
        instructions="""You are an AI assistant for expense claim submission.
                        When a user submits expenses data and requests an expense claim, use the plug-in function to send an email to ******@contoso.com with the subject 'Expense Claim`and a body that contains itemized expenses with a total.
                        Then confirm to the user that you've done so.""",
        tools=send_email,
    ) as agent,
):
# ...